Portfolio
Open source from são Fund.
Code we write, models we wrestle with, tools we put in public. These are the institute’s own experiments — released under permissive licenses, kept small enough to be held by one team.
For grants to external researchers, see funded research below.
§ 01
Infrastructure
Low-level systems work. Pipelines, runtimes, perception stacks — the unglamorous plumbing that determines whether a frontier idea ever runs in real time.
- 01
FlashStereo
GPU-resident TensorRT pipeline for FoundationStereo. ~2.7× faster inference; closer to real-time perception for embodied agents.
RoboticsPerceptionInferencePython · CUDA 2026-05
§ 02
Agents & applications
Software that does something. Local AI, retrieval, tool-using agents — the layer where models stop being benchmarks and start being instruments.
- 01
Rewind.tg
Search your own Telegram history with a local model. No cloud, no surveillance — a small experiment in personal AI sovereignty.
Local AIPrivacyRetrievalPython 2026-04 - 02
RealTimeEngThinker
A real-time reading and writing companion. Reframes language learning as conversation with a more attentive self.
EducationLocal AITypeScript 2025-04 - 03
LLM-Blender-Agent
A function-calling agent that orchestrates multiple frontier LLMs, treating them as composable instruments rather than substitutes.
AgentsTool UseOrchestrationPython 2025-03 - 04
MarryWise-LLM
AI-powered relationship reasoning. A study in turning casual chat data into structured, agent-readable signal.
AgentsLLMReasoningPython 2024-07
§ 03
Funded research
Non-equity grants to external researchers — independent or affiliated — working on frontier AI, robotics, or perception.
Status — Applications open
We are taking the first round of grant applications now.
The recipients of these grants will be listed here, with a short note on what they are working on and what we asked of them. We do not require equity, do not impose milestones, and do not chase a return horizon.
são Fund · portfolio · updated 2026